Exposing Long-Tail Safety Failures in Large Language Models through Efficient Diverse Response Sampling

Safety tuning through supervised fine-tuning and reinforcement learning from human feedback has substantially improved the robustness of large language models (LLMs). However, it often suppresses rather than eliminates unsafe behaviors, leaving rare but critical failures hidden in the long tail of the output distribution. While most red-teaming work emphasizes adversarial prompt search (input-space optimization), we show that safety failures can also be systematically exposed through diverse response generation (output-space exploration) for a fixed safety-critical prompt, where increasing the number and diversity of sampled responses can drive jailbreak success rates close to unity. To efficiently uncover such failures, we propose Progressive Diverse Population Sampling (PDPS), which combines stochastic token-level sampling with diversity-aware selection to explore a large candidate pool of responses and retain a compact, semantically diverse subset. Across multiple jailbreak benchmarks and open-source LLMs, PDPS achieves attack success rates comparable to large-scale IID sampling while using only 8% to 29% of the computational cost. Under limited-response settings, it improves success rates by 26% to 40% over IID sampling and Diverse Beam Search. Furthermore, responses generated by PDPS exhibit both a higher number and greater diversity of unsafe outputs, demonstrating its effectiveness in uncovering a broader range of failures.

From Images to Words: Efficient Cross-Modal Knowledge Distillation to Language Models from Black-box Teachers

Knowledge distillation (KD) methods are pivotal in compressing large pre-trained language models into smaller models, ensuring computational efficiency without significantly dropping performance. Traditional KD techniques assume homogeneity in modalities between the teacher (source) and the student (target) models. On the other hand, existing multimodal knowledge distillation methods require modality-specific pre-training of the teacher model, which is computationally infeasible in most cases. In this paper, we introduce ARMADA, an efficient cross-modal knowledge distillation framework designed to transfer knowledge from large vision-language models, including black-box models, to language-only models. Unlike existing KD techniques that rely on the internal structures of multimodal teachers or require computationally expensive pre-training, ARMADA leverages novel alignment techniques to distil knowledge without altering the teacher model, ensuring efficiency and scalability. We empirically validate ARMADA on twelve natural language understanding, eight complex generative reasoning and five instruction-tuning tasks, demonstrating consistent performance improvements in large models such as DeBERTa-v2-1.4B, OPT-1.3B, LLaMA-{3B, 7B, 8B}. ARMADA achieves up to 3.4% improvement on language understanding tasks and 2.6% boost in generative reasoning, all without requiring expensive multimodal pre-training or fine-tuning of the teacher model. Our findings challenge conventional knowledge distillation paradigms by demonstrating that even vision-language models, despite lacking direct textual understanding, can significantly enhance language models when distilled appropriately.

Understanding the Physics of Key-Value Cache Compression for LLMs through Attention Dynamics

As context windows in LLMs scale to 100K+ tokens, the key-value (KV) cache becomes the dominant memory bottleneck, with recent methods claiming 80-90% savings and minimal benchmark degradation. We argue these evaluations miss a structural issue: attention is not just storage but routing, and retaining KV pairs does not guarantee semantic accessibility. We propose a physics-inspired view of KV compression as a controlled perturbation of token-level routing, distinguishing retention, accessibility, and utilization. Using synthetic tasks probing multi-entity tracking, disambiguation, coreference, and multi-hop reasoning, we find that moderate compression degrades internal representations with little accuracy loss, revealing redundancy; all models exhibit a sharp hallucination safety cliff near 90% compression, correlated with spikes in Global Eviction Ratio (GER), suggesting a phase transition in semantic reachability; and architectures differ in routing dynamics, with LLaMA showing early consensus and late diversification, and Qwen showing funnel-like late convergence, leading to distinct resilience profiles. Beyond erasure, we identify representational rigidity, where excessive head-level consensus collapses routing flexibility despite token survival. These results suggest sparse token-route structures govern compression tolerance, reframing KV compression as a structural probe of attention geometry and linking long-context scalability to sparsity and the lottery ticket hypothesis in self-attention.

Markovian ODE-guided scoring can assess the quality of offline reasoning traces in language models

Reasoning traces produced by generative language models are increasingly used for tasks ranging from mathematical problem solving to automated fact checking. However, existing evaluation methods remain largely mechanical and fail to capture human-centric notions of reasoning quality in a way that generalizes across varied and progressively degraded reasoning. We introduce MarODE, an offline evaluation framework that assigns quality scores to reasoning traces. Its effectiveness is assessed using human-centric perturbations and human judgments, which jointly evaluate the fundamental dimensions of an evaluation metric - goodness and soundness. The approach is grounded in a Markovian formulation of reasoning progression and an ordinary differential equation based characterization of trace dynamics, enabling efficient evaluation of reasoning quality. In a large-scale evaluation, MarODE outperforms existing baselines by over 250% under Somers' D correlation. Our results emphasize the value of theory-driven evaluation frameworks as reasoning traces become central to language model-based systems.

The CLEF-2026 CheckThat! Lab: Advancing Multilingual Fact-Checking

The CheckThat! lab aims to advance the development of innovative technologies combating disinformation and manipulation efforts in online communication across a multitude of languages and platforms. While in early editions the focus has been on core tasks of the verification pipeline (check-worthiness, evidence retrieval, and verification), in the past three editions, the lab added additional tasks linked to the verification process. In this year's edition, the verification pipeline is at the center again with the following tasks: Task 1 on source retrieval for scientific web claims (a follow-up of the 2025 edition), Task 2 on fact-checking numerical and temporal claims, which adds a reasoning component to the 2025 edition, and Task 3, which expands the verification pipeline with generation of full-fact-checking articles. These tasks represent challenging classification and retrieval problems as well as generation challenges at the document and span level, including multilingual settings.

Linguistic properties and model scale in brain encoding: from small to compressed language models

Recent work has shown that scaling large language models (LLMs) improves their alignment with human brain activity, yet it remains unclear what drives these gains and which representational properties are responsible. Although larger models often yield better task performance and brain alignment, they are increasingly difficult to analyze mechanistically. This raises a fundamental question: what is the minimal model capacity required to capture brain-relevant representations? To address this question, we systematically investigate how constraining model scale and numerical precision affects brain alignment. We compare full-precision LLMs, small language models (SLMs), and compressed variants (quantized and pruned) by predicting fMRI responses during naturalistic language comprehension. Across model families up to 14B parameters, we find that 3B SLMs achieve brain predictivity indistinguishable from larger LLMs, whereas 1B models degrade substantially, particularly in semantic language regions. Brain alignment is remarkably robust to compression: most quantization and pruning methods preserve neural predictivity, with GPTQ as a consistent exception. Linguistic probing reveals a dissociation between task performance and brain predictivity: compression degrades discourse, syntax, and morphology, yet brain predictivity remains largely unchanged. Overall, brain alignment saturates at modest model scales and is resilient to compression, challenging common assumptions about neural scaling and motivating compact models for brain-aligned language modeling.

How does longer temporal context enhance multimodal narrative video processing in the brain?

Understanding how humans and artificial intelligence systems process complex narrative videos is a fundamental challenge at the intersection of neuroscience and machine learning. This study investigates how the temporal context length of video clips (3--12 s clips) and the narrative-task prompting shape brain-model alignment during naturalistic movie watching. Using fMRI recordings from participants viewing full-length movies, we examine how brain regions sensitive to narrative context dynamically represent information over varying timescales and how these neural patterns align with model-derived features. We find that increasing clip duration substantially improves brain alignment for multimodal large language models (MLLMs), whereas unimodal video models show little to no gain. Further, shorter temporal windows align with perceptual and early language regions, while longer windows preferentially align higher-order integrative regions, mirrored by a layer-to-cortex hierarchy in MLLMs. Finally, narrative-task prompts (multi-scene summary, narrative summary, character motivation, and event boundary detection) elicit task-specific, region-dependent brain alignment patterns and context-dependent shifts in clip-level tuning in higher-order regions. Together, our results position long-form narrative movies as a principled testbed for probing biologically relevant temporal integration and interpretable representations in long-context MLLMs.

Disentangling Causal Importance from Emergent Structure in Multi-Expert Orchestration

Multi-expert systems, where multiple Large Language Models (LLMs) collaborate to solve complex tasks, are increasingly adopted for high-performance reasoning and generation. However, the orchestration policies governing expert interaction and sequencing remain largely opaque. We introduce INFORM, an interpretability analysis that treats orchestration as an explicit, analyzable computation, enabling the decoupling of expert interaction structure, execution order, and causal attribution. We use INFORM to evaluate an orchestrator on GSM8K, HumanEval, and MMLU using a homogeneous consortium of ten instruction-tuned experts drawn from LLaMA-3.1 8B, Qwen-3 8B, and DeepSeek-R1 8B, with controlled decoding-temperature variation, and a secondary heterogeneous consortium spanning 1B-7B parameter models. Across tasks, routing dominance is a poor proxy for functional necessity. We reveal a divergence between relational importance, captured by routing mass and interaction topology, and intrinsic importance, measured via gradient-based causal attribution: frequently selected experts often act as interaction hubs with limited causal influence, while sparsely routed experts can be structurally critical. Orchestration behaviors emerge asynchronously, with expert centralization preceding stable routing confidence and expert ordering remaining non-deterministic. Targeted ablations show that masking intrinsically important experts induces disproportionate collapse in interaction structure compared to masking frequent peers, confirming that INFORM exposes causal and structural dependencies beyond accuracy metrics alone.

SpatialMath: Spatial Comprehension-Infused Symbolic Reasoning for Mathematical Problem-Solving

Multimodal Small-to-Medium sized Language Models (MSLMs) have demonstrated strong capabilities in integrating visual and textual information but still face significant limitations in visual comprehension and mathematical reasoning, particularly in geometric problems with diverse levels of visual infusion. Current models struggle to accurately decompose intricate visual inputs and connect perception with structured reasoning, leading to suboptimal performance. To address these challenges, we propose SpatialMath, a novel Spatial Comprehension-Infused Symbolic Reasoning Framework designed to integrate spatial representations into structured symbolic reasoning chains. SpatialMath employs a specialized perception module to extract spatially-grounded representations from visual diagrams, capturing critical geometric structures and spatial relationships. These representations are then methodically infused into symbolic reasoning chains, facilitating visual comprehension-aware structured reasoning. To this end, we introduce MATHVERSE-PLUS, a novel dataset containing structured visual interpretations and step-by-step reasoning paths for vision-intensive mathematical problems. SpatialMath significantly outperforms strong multimodal baselines, achieving up to 10 percentage points improvement over supervised fine-tuning with data augmentation in vision-intensive settings. Robustness analysis reveals that enhanced spatial representations directly improve reasoning accuracy, reinforcing the need for structured perception-to-reasoning pipelines in MSLMs.

coTherapist: A Behavior-Aligned Small Language Model to Support Mental Healthcare Experts

Access to mental healthcare is increasingly strained by workforce shortages and rising demand, motivating the development of intelligent systems that can support mental healthcare experts. We introduce coTherapist, a unified framework utilizing a small language model to emulate core therapeutic competencies through domain-specific fine-tuning, retrieval augmentation, and agentic reasoning. Evaluation on clinical queries demonstrates that coTherapist generates more relevant and clinically grounded responses than contemporary baselines. Using our novel T-BARS rubric and psychometric profiling, we confirm coTherapist exhibits high empathy and therapist-consistent personality traits. Furthermore, human evaluation by domain experts validates that coTherapist delivers accurate, trustworthy, and safe responses. coTherapist was deployed and tested by clinical experts. Collectively, these findings demonstrate that small models can be engineered to exhibit expert-like behavior, offering a scalable pathway for digital mental health tools.